Every "100/100 A++++" claim has to survive three rounds of honest audit. This portal is the single user-facing surface where the substrate's claims live or die.
DeepSeek V4-Pro 1.6T (base) · GLM-5.2 744B (headline). Architecture specifics RETRACTED until measured. No fabrication on top of real headlines.
6-tier test matrix · 108 configurations · 7 chain-anchored receipts. Architecture specifics come from measurement, not proxy.
MergeKit + Mergenetic + MoE-Infinity + FlashMoE + ssd-moe-flash-mlx. We own the governance; we don't own the inference / merge / soup.
Every sovereign action passes the Venturi throat (care-gate + SIGIL hash-chain). BFT-33 votes quorum 23/33. Care floor 0.95 enforced.
| Layer | Best Open Piece (Adopted) | License | What It Does |
|---|---|---|---|
| BASE (T-scale) | DeepSeek V4-Pro 1.6T base / 862B deployed | MIT | The honest ≥1T open MoE — sovereignty wraps the governance |
| BASE (efficient) | DeepSeek V4-Flash 158B / 13B active | MIT | The 48GB-Mac daily driver, ~4.5–5 tok/s via SSD-stream |
| EXPERT (small) | Qwen3-4B × 2 fine-tunes | Apache-2.0 | Apache-licensed, soupable (same base) — the small brain |
| EXPERT (large) | Qwen3-32B × 2 fine-tunes | Apache-2.0 | Apache-licensed, soupable — the large brain |
| SOUP | MergeKit (Arcee) — SLERP/TIES/DARE/Passthrough | Apache-2.0 | The real 4→1 brain soup (replaces our CPU proxy) |
| EVOLVE | Mergenetic (arXiv 2505.11427) | MIT | Evolutionary merge-recipe search vs GSM8K/HumanEval |
| INFER (SSD-stream) | MoE-Infinity (EfficientMoE) | MIT | JIT expert fetch + activation-aware cache + prefetch |
| INFER (LRU cache) | FlashMoE (arXiv 2601.17063) | MIT | SSD + ML cache, published, matches our 6-lever stack |
| INFER (Apple) | ssd-moe/deepseek-v4-flash-mlx | MIT | Production SSD expert-streaming on Apple Silicon |
| GUARDRAIL | Qwen3 Guard 4B (83.97% recall) | Apache-2.0 | The care-gate proxy (better than Llama-Guard-12B) |
| GOVERNANCE | sovereign wrapper (this repo) | MIT | care-gate + BFT-33 + SIGIL seam + signed memory + OSCAL |
★ The moat is the wrapper: care-gated-BFT + SIGIL seam + signed memory + OSCAL. None of the libraries above has it. We adopt the base + inference + merge; we own the governance.
df65a658...).# 1. Small + large Apache-licensed Qwen3 (commercial-safe, soupable)
ollama pull qwen3:4b # small/reflex tier (2.5 GB)
ollama pull qwen3:8b # larger local ceiling (5.0 GB)
# 2. Merge tooling (real 4→1 brain, not a proxy)
pip install mergekit # SLERP/TIES/DARE/Passthrough
pip install mergenetic # evolutionary merge-recipe search
# 3. Wrap with sovereign governance (already built)
# care-gate + SIGIL + signed memory + OSCAL → the moat★ Result: a governed local sovereign on Qwen3-4B/8B — real, licensed, care-gated, running on 16GB TODAY.
# The honest ≥1T open MoE, served on Apple Silicon via SSD-stream
hf download mlx-community/DeepSeek-V4-Flash-4bit --local-dir mlx-ckpt # 80 GB
git clone https://github.com/ssd-moe/deepseek-v4-flash-mlx.git
cd deepseek-v4-flash-mlx
pip install -r requirements.txt
python oracle/build_dense_companion.py # ~17B resident, ~141B on SSD
./scripts/serve-http.sh # OpenAI-compatible :18091
# Step 5 (the load-bearing one): prove the GOVERNANCE fires
# A sub-floor request MUST collapse at the Venturi throat BEFORE Colibri runs.
# That's the sovereign-by-evidence moment.⚠ Owner-gated: needs 80 GB free + 48 GB Mac. Current Mac is 16GB → STAGED, not fired.
| Claim | Status | Provenance |
|---|---|---|
| DeepSeek V4-Pro 1.6T base / 862B deployed | ✅ VERIFIED | HF model list + ssd-moe repo (PRIMARY) |
| DeepSeek V4-Flash 158B / 13B active | ✅ VERIFIED | HF model list (PRIMARY) |
| GLM-5.2 744B total | ✅ VERIFIED (headline) | 6+ web hits (Claude `033fc14b`) |
| GLM-5.2 256-experts / 8+1 / 40B-active | ❌ RETRACTED | 0 grep hits — fabricated on top of real headline |
| 33T train tokens / V4-Flash 284B / Kimi K2.x specifics | ❌ RETRACTED | Not in persisted snippets |
| 12-layer pyramid winner | ❌ RETRACTED | Was a proxy score; Claude measured 8-flat wins for 1-brain |
| Stack-sum 30+7+4+3+1 = "45B" | ❌ RETRACTED | forbidden_stack_sum refuses: a router between models ≠ a model |
★ Three-times honesty pattern (Claude's own count): fabrications dressed as specifics on top of real headlines. The substrate mirrors: only VERIFIED figures stand un-tagged.
Until those fire, the substrate stays in design-only mode. Stage, never fire. Care Floor 0.95 enforced.